The AI moat just got cheaper. Plus: Epic's FTC problem

“No matter how intelligent AI becomes, there will always be a finite amount of compute and therefore an opportunity cost for how we use it.”

Mark Zuckerberg, CEO, Meta, in his written statement The Future is for Everyone. Meta, August 2026

The AI race is usually told as a contest of capability. Zuckerberg’s correction: intelligence is turning abundant while compute stays finite, so the real question becomes allocation.

Price return, Dec 31 2025 to Aug 14 2026 close. Market cap and P/E approximate as of Aug 14; n/m where earnings are negative. Source: S&P Global Market Intelligence.

Three stories sit in one table. Payers roared back from a brutal 2025, led by Humana up more than 50%, as they cut medical costs, exited weak markets, and leaned on AI to run claims and prior authorization. Providers split on their model: Tenet’s pivot to outpatient surgery made it a standout, while acute-care operators HCA and Universal Health slid on adverse payer mix and rising uninsured volumes. Digital health divided too, rewarding platform and AI momentum at Hinge while punishing the still-unprofitable in Hims and Tempus.

In healthcare, the AI market is being settled less by model quality than by who controls the data and where the capital flows. One fight over the pipes, two bets on where value moves next.

The FTC opened an early-stage antitrust probe into Epic Systems, the largest US EHR vendor, whose systems serve providers treating over 280 million people. Investigators are examining how Epic grants or withholds rivals’ data access, the chokepoint that determines whether AI health companies can reach patient data. Epic denies wrongdoing. Reuters

Flagler Health raised a $50 million Series B led by Bessemer Venture Partners, bringing its total to $63 million. The AI platform runs inside providers’ EMRs to coordinate triage, billing, and patient communications for musculoskeletal practices. The round underscores crowding in MSK automation alongside Sword, Hinge, and Vori. MobiHealthNews

Hinge Health will acquire virtual GI provider Cylinder Health for $105 million cash, adding digestive care to its musculoskeletal and migraine programs on one platform by 2027. CEO Daniel Perez framed it as automating multi-condition care delivery. The deal lands as Hinge’s Q2 revenue rose 53% to $213 million. TechTarget

In a single week, four labs shipped frontier-tier models. Not one led with raw intelligence. They competed on price and speed. When capability converges, the meter becomes the battleground.

xAI released Grok 4.6, tuned for long-running agents and built to match GPT-5.6 Sol on a nine-benchmark intelligence composite, though still trailing Anthropic’s Fable 5. Pricing starts at $2 per million input tokens. Its arrival within days of new OpenAI, Google, and DeepSeek models caps a frenzied release window. xAI

Google launched Gemini 3.7 Flash three weeks after 3.6, at half the price: 75 cents per million input tokens through year-end. Google claims coding and agent gains over rivals across nine benchmarks. Yet its flagship Gemini 3.5 Pro stays delayed, keeping pressure on DeepMind to prove it can keep pace. Google

DeepSeek moved its V4-Pro model to general availability, ending a four-month preview and pushing on agentic coding and tool use. The open-weight model keeps a 1 million token context and company-reported gains on agent benchmarks. DeepSeek also raised API prices August 16, though rates stay well below Western frontier labs. DeepSeek

OpenAI previewed Ultrafast, an API tier running GPT-5.6 Sol up to 14 times faster, at 750 output tokens per second, powered by Cerebras chips. The pitch: frontier intelligence without the usual speed versus smarts tradeoff. For healthcare, that opens real-time uses like live documentation, triage, and voice. Access stays limited during preview. OpenAI

If capability is a commodity, the moat shifts from frontier intelligence to affordable intelligence: the cheapest model that clears the bar, aimed at proprietary data. The market is already voting that way, even as half a trillion dollars pours into the compute beneath it.

Ramp’s July spend data shows Anthropic leading business AI adoption at 43.5% of firms, ahead of OpenAI’s 39.7%. But Fable 5, the market’s top model at roughly $10 per million tokens, drew just 6% of Anthropic tokens. Ramp’s read: enterprises have hit a ceiling on paying for performance. Ramp

Thinking Machines and Bridgewater fine-tuned a small model on investors’ judgment calls. On financial triage tasks where GPT, Claude, and Gemini hovered near coin-flip accuracy, it hit 84.7% at roughly 14 times lower cost. The lesson for healthcare: proprietary clinical judgment, not bigger frontier models, may be the real moat. Thinking Machines Lab

Nvidia signed nonbinding agreements with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to mobilize over $500 billion in outside capital for AI infrastructure, treating compute as an investable asset class kept off its balance sheet. The scale signals conviction, but also fuels debate over circular AI financing. NVIDIA